Reimplement rc4-586.pl, relicense rc4-x86_64.pl.
authorAndy Polyakov <appro@openssl.org>
Thu, 26 Apr 2007 20:48:38 +0000 (20:48 +0000)
committerAndy Polyakov <appro@openssl.org>
Thu, 26 Apr 2007 20:48:38 +0000 (20:48 +0000)
crypto/rc4/asm/rc4-586.pl
crypto/rc4/asm/rc4-x86_64.pl

index 9766abf..ef148b2 100644 (file)
@@ -1,14 +1,21 @@
-#!/usr/local/bin/perl
+#!/usr/bin/env perl
+
+# ====================================================================
+# [Re]written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
+# project. The module is, however, dual licensed under OpenSSL and
+# CRYPTOGAMS licenses depending on where you obtain it. For further
+# details see http://www.openssl.org/~appro/cryptogams/.
+# ====================================================================
 
 # At some point it became apparent that the original SSLeay RC4
-# assembler implementation performs suboptimaly on latest IA-32
+# assembler implementation performs suboptimally on latest IA-32
 # microarchitectures. After re-tuning performance has changed as
 # following:
 #
-# Pentium      +0%
-# Pentium III  +17%
-# AMD          +52%(*)
-# P4           +180%(**)
+# Pentium      -10%
+# Pentium III  +12%
+# AMD          +50%(*)
+# P4           +250%(**)
 #
 # (*)  This number is actually a trade-off:-) It's possible to
 #      achieve +72%, but at the cost of -48% off PIII performance.
 #      For reference! This code delivers ~80% of rc4-amd64.pl
 #      performance on the same Opteron machine.
 # (**) This number requires compressed key schedule set up by
-#      RC4_set_key and therefore doesn't apply to 0.9.7 [option for
-#      compressed key schedule is implemented in 0.9.8 and later,
-#      see commentary section in rc4_skey.c for further details].
+#      RC4_set_key [see commentary below for further details].
 #
 #                                      <appro@fy.chalmers.se>
 
-push(@INC,"perlasm","../../perlasm");
+$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
+push(@INC,"${dir}","${dir}../../perlasm");
 require "x86asm.pl";
 
 &asm_init($ARGV[0],"rc4-586.pl");
 
-$x="eax";
-$y="ebx";
+$xx="eax";
+$yy="ebx";
 $tx="ecx";
 $ty="edx";
-$in="esi";
-$out="edi";
-$d="ebp";
-
-sub RC4_loop
-       {
-       local($n,$p,$char)=@_;
-
-       &comment("Round $n");
-
-       if ($char)
-               {
-               if ($p >= 0)
-                       {
-                        &mov($ty,      &swtmp(2));
-                       &cmp($ty,       $in);
-                        &jbe(&label("finished"));
-                       &inc($in);
-                       }
-               else
-                       {
-                       &add($ty,       8);
-                        &inc($in);
-                       &cmp($ty,       $in);
-                        &jb(&label("finished"));
-                       &mov(&swtmp(2), $ty);
-                       }
-               }
-       # Moved out
-       # &mov( $tx,            &DWP(0,$d,$x,4)) if $p < 0;
-
-       &add(   &LB($y),        &LB($tx));
-       &mov(   $ty,            &DWP(0,$d,$y,4));
-        # XXX
-       &mov(   &DWP(0,$d,$x,4),$ty);
-        &add(  $ty,            $tx);
-       &mov(   &DWP(0,$d,$y,4),$tx);
-        &and(  $ty,            0xff);
-        &inc(  &LB($x));                       # NEXT ROUND
-       &mov(   $tx,            &DWP(0,$d,$x,4)) if $p < 1; # NEXT ROUND
-        &mov(  $ty,            &DWP(0,$d,$ty,4));
-
-       if (!$char)
-               {
-               #moved up into last round
-               if ($p >= 1)
-                       {
-                       &add(   $out,   8)
-                       }
-               &movb(  &BP($n,"esp","",0),     &LB($ty));
-               }
-       else
-               {
-               # Note in+=8 has occured
-               &movb(  &HB($ty),       &BP(-1,$in,"",0));
-                # XXX
-               &xorb(&LB($ty),         &HB($ty));
-                # XXX
-               &movb(&BP($n,$out,"",0),&LB($ty));
-               }
+$inp="esi";
+$out="ebp";
+$dat="edi";
+
+sub RC4_loop {
+  my $i=shift;
+  my $func = ($i==0)?*mov:*or;
+
+       &add    (&LB($yy),&LB($tx));
+       &mov    ($ty,&DWP(0,$dat,$yy,4));
+       &mov    (&DWP(0,$dat,$yy,4),$tx);
+       &mov    (&DWP(0,$dat,$xx,4),$ty);
+       &add    ($ty,$tx);
+       &inc    (&LB($xx));
+       &and    ($ty,0xff);
+       &ror    ($out,8)        if ($i!=0);
+       if ($i<3) {
+         &mov  ($tx,&DWP(0,$dat,$xx,4));
+       } else {
+         &mov  ($tx,&wparam(3));       # reload [re-biased] out
        }
+       &$func  ($out,&DWP(0,$dat,$ty,4));
+}
 
-
-&function_begin_B("RC4");
-       {
-       local($name)=@_;
-
-
-       &mov($ty,&wparam(1));           # len
-       &cmp($ty,0);
-       &jne(&label("proceed"));
-       &ret();
-       &set_label("proceed");
-
-       &comment("");
-
-       &push("ebp");
-        &push("ebx");
-       &push("esi");
-        &xor(  $x,     $x);            # avoid partial register stalls
-       &push("edi");
-        &xor(  $y,     $y);            # avoid partial register stalls
-       &mov(   $d,     &wparam(0));    # key
-        &mov(  $in,    &wparam(2));
-
-       &movb(  &LB($x),        &BP(0,$d,"",1));
-        &movb( &LB($y),        &BP(4,$d,"",1));
-
-       &mov(   $out,   &wparam(3));
-        &inc(  &LB($x));
-
-       &stack_push(3); # 3 temp variables
-        &add(  $d,     8);
-
-       # detect compressed schedule, see commentary section in rc4_skey.c...
-       # in 0.9.7 context ~50 bytes below RC4_CHAR label remain redundant,
-       # as compressed key schedule is set up in 0.9.8 and later.
-       &cmp(&DWP(256,$d),-1);
-       &je(&label("RC4_CHAR"));
-
-        &lea(  $ty,    &DWP(-8,$ty,$in));
-
-       # check for 0 length input
-
-        &mov(  &swtmp(2),      $ty);   # this is now address to exit at
-       &mov(   $tx,    &DWP(0,$d,$x,4));
-
-        &cmp(  $ty,    $in);
-       &jb(    &label("end")); # less than 8 bytes
-
-       &set_label("start");
-
-       # filling DELAY SLOT
-       &add(   $in,    8);
-
-       &RC4_loop(0,-1,0);
-       &RC4_loop(1,0,0);
-       &RC4_loop(2,0,0);
-       &RC4_loop(3,0,0);
-       &RC4_loop(4,0,0);
-       &RC4_loop(5,0,0);
-       &RC4_loop(6,0,0);
-       &RC4_loop(7,1,0);
-       
-       &comment("apply the cipher text");
-       # xor the cipher data with input
-
-       #&add(  $out,   8); #moved up into last round
-
-       &mov(   $tx,    &swtmp(0));
-        &mov(  $ty,    &DWP(-8,$in,"",0));
-       &xor(   $tx,    $ty);
-        &mov(  $ty,    &DWP(-4,$in,"",0)); 
-       &mov(   &DWP(-8,$out,"",0),     $tx);
-        &mov(  $tx,    &swtmp(1));
-       &xor(   $tx,    $ty);
-        &mov(  $ty,    &swtmp(2));     # load end ptr;
-       &mov(   &DWP(-4,$out,"",0),     $tx);
-        &mov(  $tx,            &DWP(0,$d,$x,4));
-       &cmp($in,       $ty);
-        &jbe(&label("start"));
-
-       &set_label("end");
-
-       # There is quite a bit of extra crap in RC4_loop() for this
-       # first round
-       &RC4_loop(0,-1,1);
-       &RC4_loop(1,0,1);
-       &RC4_loop(2,0,1);
-       &RC4_loop(3,0,1);
-       &RC4_loop(4,0,1);
-       &RC4_loop(5,0,1);
-       &RC4_loop(6,1,1);
-
-       &jmp(&label("finished"));
-
-       &align(16);
-       # this is essentially Intel P4 specific codepath, see rc4_skey.c,
-       # and is engaged in 0.9.8 and later context...
-       &set_label("RC4_CHAR");
-
-       &lea    ($ty,&DWP(0,$in,$ty));
-       &mov    (&swtmp(2),$ty);
-       &movz   ($tx,&BP(0,$d,$x));
-
+# void RC4(RC4_KEY *key,size_t len,const unsigned char *inp,unsigned char *out);
+&function_begin("RC4");
+       &mov    ($dat,&wparam(0));      # load key schedule pointer
+       &mov    ($ty, &wparam(1));      # load len
+       &mov    ($inp,&wparam(2));      # load inp
+       &mov    ($out,&wparam(3));      # load out
+
+       &xor    ($xx,$xx);              # avoid partial register stalls
+       &xor    ($yy,$yy);
+
+       &cmp    ($ty,0);                # safety net
+       &je     (&label("abort"));
+
+       &mov    (&LB($xx),&BP(0,$dat)); # load key->x
+       &mov    (&LB($yy),&BP(4,$dat)); # load key->y
+       &add    ($dat,8);
+
+       &lea    ($tx,&DWP(0,$inp,$ty));
+       &sub    ($out,$inp);            # re-bias out
+       &mov    (&wparam(1),$tx);       # save input+len
+
+       &inc    (&LB($xx));
+
+       # detect compressed key schedule...
+       &cmp    (&DWP(256,$dat),-1);
+       &je     (&label("RC4_CHAR"));
+
+       &mov    ($tx,&DWP(0,$dat,$xx,4));
+
+       &and    ($ty,-4);               # how many 4-byte chunks?
+       &jz     (&label("loop1"));
+
+       &lea    ($ty,&DWP(-4,$inp,$ty));
+       &mov    (&wparam(2),$ty);       # save input+(len/4)*4-4
+       &mov    (&wparam(3),$out);      # $out as accumulator in this loop
+
+       &set_label("loop4",16);
+               for ($i=0;$i<4;$i++) { RC4_loop($i); }
+               &ror    ($out,8);
+               &xor    ($out,&DWP(0,$inp));
+               &cmp    ($inp,&wparam(2));      # compare to input+(len/4)*4-4
+               &mov    (&DWP(0,$tx,$inp),$out);# $tx holds re-biased out here
+               &lea    ($inp,&DWP(4,$inp));
+               &mov    ($tx,&DWP(0,$dat,$xx,4));
+       &jb     (&label("loop4"));
+
+       &cmp    ($inp,&wparam(1));      # compare to input+len
+       &je     (&label("done"));
+       &mov    ($out,&wparam(3));      # restore $out
+
+       &set_label("loop1",16);
+               &add    (&LB($yy),&LB($tx));
+               &mov    ($ty,&DWP(0,$dat,$yy,4));
+               &mov    (&DWP(0,$dat,$yy,4),$tx);
+               &mov    (&DWP(0,$dat,$xx,4),$ty);
+               &add    ($ty,$tx);
+               &inc    (&LB($xx));
+               &and    ($ty,0xff);
+               &mov    ($ty,&DWP(0,$dat,$ty,4));
+               &xor    (&LB($ty),&BP(0,$inp));
+               &lea    ($inp,&DWP(1,$inp));
+               &mov    ($tx,&DWP(0,$dat,$xx,4));
+               &cmp    ($inp,&wparam(1));      # compare to input+len
+               &mov    (&BP(-1,$out,$inp),&LB($ty));
+       &jb     (&label("loop1"));
+
+       &jmp    (&label("done"));
+
+# this is essentially Intel P4 specific codepath...
+&set_label("RC4_CHAR",16);
+       &movz   ($tx,&BP(0,$dat,$xx));
        # strangely enough unrolled loop performs over 20% slower...
-       &set_label("RC4_CHAR_loop");
-               &add    (&LB($y),&LB($tx));
-               &movz   ($ty,&BP(0,$d,$y));
-               &movb   (&BP(0,$d,$y),&LB($tx));
-               &movb   (&BP(0,$d,$x),&LB($ty));
+       &set_label("cloop1");
+               &add    (&LB($yy),&LB($tx));
+               &movz   ($ty,&BP(0,$dat,$yy));
+               &mov    (&BP(0,$dat,$yy),&LB($tx));
+               &mov    (&BP(0,$dat,$xx),&LB($ty));
                &add    (&LB($ty),&LB($tx));
-               &movz   ($ty,&BP(0,$d,$ty));
-               &add    (&LB($x),1);
-               &xorb   (&LB($ty),&BP(0,$in));
-               &lea    ($in,&BP(1,$in));
-               &movz   ($tx,&BP(0,$d,$x));
-               &cmp    ($in,&swtmp(2));
-               &movb   (&BP(0,$out),&LB($ty));
-               &lea    ($out,&BP(1,$out));
-       &jb     (&label("RC4_CHAR_loop"));
-
-       &set_label("finished");
-       &dec(   $x);
-        &stack_pop(3);
-       &movb(  &BP(-4,$d,"",0),&LB($y));
-        &movb( &BP(-8,$d,"",0),&LB($x));
-}
+               &movz   ($ty,&BP(0,$dat,$ty));
+               &add    (&LB($xx),1);
+               &xor    (&LB($ty),&BP(0,$inp));
+               &lea    ($inp,&BP(1,$inp));
+               &movz   ($tx,&BP(0,$dat,$xx));
+               &cmp    ($inp,&wparam(1));
+               &mov    (&BP(-1,$out,$inp),&LB($ty));
+       &jb     (&label("cloop1"));
+
+&set_label("done");
+       &dec    (&LB($xx));
+       &mov    (&BP(-4,$dat),&LB($yy));        # save key->y
+       &mov    (&BP(-8,$dat),&LB($xx));        # save key->x
+&set_label("abort");
 &function_end("RC4");
 
 ########################################################################
@@ -271,6 +205,17 @@ $idx="edx";
        &jnc    (&label("w2ndloop"));
 &jmp   (&label("exit"));
 
+# Unlike all other x86 [and x86_64] implementations, Intel P4 core
+# [including EM64T] was found to perform poorly with above "32-bit" key
+# schedule, a.k.a. RC4_INT. Performance improvement for IA-32 hand-coded
+# assembler turned out to be 3.5x if re-coded for compressed 8-bit one,
+# a.k.a. RC4_CHAR! It's however inappropriate to just switch to 8-bit
+# schedule for x86[_64], because non-P4 implementations suffer from
+# significant performance losses then, e.g. PIII exhibits >2x
+# deterioration, and so does Opteron. In order to assure optimal
+# all-round performance, we detect P4 at run-time and set up compressed
+# key schedule, which is recognized by RC4 procedure.
+
 &set_label("c1stloop",16);
        &mov    (&BP(0,$out,"eax"),&LB("eax")); # key->data[i]=i;
        &add    (&LB("eax"),1);                 # i++;
@@ -315,9 +260,9 @@ $idx="edx";
        &set_label("skip");
        &ret    ();
 &set_label("opts",64);
-&asciz ("rc4(8x,int)");
+&asciz ("rc4(4x,int)");
 &asciz ("rc4(1x,char)");
-&asciz ("RC4 for x86, OpenSSL project");       # RC4_version
+&asciz ("RC4 for x86, CRYPTOGAMS by <appro\@openssl.org>");
 &align (64);
 &function_end_B("RC4_options");
 
index 36a9429..2d47320 100755 (executable)
@@ -2,8 +2,9 @@
 #
 # ====================================================================
 # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
-# project. Rights for redistribution and usage in source and binary
-# forms are granted according to the OpenSSL license.
+# project. The module is, however, dual licensed under OpenSSL and
+# CRYPTOGAMS licenses depending on where you obtain it. For further
+# details see http://www.openssl.org/~appro/cryptogams/.
 # ====================================================================
 #
 # 2.22x RC4 tune-up:-) It should be noted though that my hand [as in
 # this CPU.
 
 $output=shift;
-open STDOUT,"| $^X ../perlasm/x86_64-xlate.pl $output";
+
+$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
+( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
+( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
+die "can't locate x86_64-xlate.pl";
+
+open STDOUT,"| $^X $xlate $output";
 
 $dat="%rdi";       # arg1
 $len="%rsi";       # arg2
@@ -345,7 +352,7 @@ RC4_options:
 .asciz "rc4(8x,int)"
 .asciz "rc4(8x,char)"
 .asciz "rc4(1x,char)"
-.asciz "RC4 for x86_64, OpenSSL project"
+.asciz "RC4 for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
 .align 64
 .size  RC4_options,.-RC4_options
 ___